Deadline Miss Rates of Applications with Stochastic Task Execution Times Sorin Manolache, Petru Eles, Zebo Peng {sorma, petel, zebpe}@ida.liu.se Department of Computer and Information Science Linköping University, Sweden Probability Motivation Expensive hardware 0% missed deadlines Probability Task execution time Affordable hardware <5% missed deadlines Task execution time Deadline Miss Rate Analysis of Applications with Stochastic Task Execution Times Sorin Manolache, Petru Eles, Zebo Peng – Linkoping University, Sweden 2 Problem formulation, input 2s graphs Task periods Task execution time probability 2s density functions 4s Message transmission time probability density functions Task and task graph deadlines 6s miss 4% Mapping of tasks to processors and messages to buses 10s Deadline miss ratio thresholds Probability Task 10s Task execution time miss 2% miss 10% Deadline Miss Rate Analysis of Applications with Stochastic Task Execution Times Sorin Manolache, Petru Eles, Zebo Peng – Linkoping University, Sweden 3 Problem formulation, output miss 0% miss 2% Deadline miss ratios per task and task graph miss 0% miss 2% miss 5% miss 0% miss 7% miss 3% Deadline Miss Rate Analysis of Applications with Stochastic Task Execution Times Sorin Manolache, Petru Eles, Zebo Peng – Linkoping University, Sweden miss 3% 4 Outline For monoprocessor systems, we found an exact solution based on concurrent construction and analysis of the underlying generalized semi-Markov process [Manolache et al. “Memory and Time-Efficient Schedulability Analysis of Task Graphs with Stochastic Execution Time”, ECRTS-01] The solution is theoretically applicable to multiprocessor systems, but practically to only very small ones, because of complexity Solutions based on approximation 1. Execution time PDF approximation 2. Independence assumption among various random variables Deadline Miss Rate Analysis of Applications with Stochastic Task Execution Times Sorin Manolache, Petru Eles, Zebo Peng – Linkoping University, Sweden 5 Execution Time PDF Approximation Coxian approximation-based Task graphs Modelling Approximation GSPN Coxian distribs CTMC constr. CTMC Results Analysis Deadline Miss Rate Analysis of Applications with Stochastic Task Execution Times Sorin Manolache, Petru Eles, Zebo Peng – Linkoping University, Sweden 7 Application modelling (1) A E B C F D Deadline Miss Rate Analysis of Applications with Stochastic Task Execution Times Sorin Manolache, Petru Eles, Zebo Peng – Linkoping University, Sweden 8 Application modelling (2) A B C D F C F probab A E D B E Firing delay equals execution time firing delay Deadline Miss Rate Analysis of Applications with Stochastic Task Execution Times Sorin Manolache, Petru Eles, Zebo Peng – Linkoping University, Sweden 9 Coxian approximation-based Task graphs Modelling Approximation GSPN Coxian distribs CTMC constr. CTMC Results Analysis Deadline Miss Rate Analysis of Applications with Stochastic Task Execution Times Sorin Manolache, Petru Eles, Zebo Peng – Linkoping University, Sweden 10 CTMC construction (1) X, Y X, Y X GSMP Approximation of the GSMP X Approximation of X Deadline Miss Rate Analysis of Applications with Stochastic Task Execution Times Sorin Manolache, Petru Eles, Zebo Peng – Linkoping University, Sweden 11 CTMC construction (2) The global generator of the Markov chain becomes then M ( Aj ) I jen ( I j ) Bi ( I j ) D j ien j i , jen j i , jen M is expressed in terms of small matrices and can be generated on the fly – memory savings Deadline Miss Rate Analysis of Applications with Stochastic Task Execution Times Sorin Manolache, Petru Eles, Zebo Peng – Linkoping University, Sweden 12 Analysis time vs. number of tasks Deadline Miss Rate Analysis of Applications with Stochastic Task Execution Times Sorin Manolache, Petru Eles, Zebo Peng – Linkoping University, Sweden 13 Analysis time vs. number of procs Deadline Miss Rate Analysis of Applications with Stochastic Task Execution Times Sorin Manolache, Petru Eles, Zebo Peng – Linkoping University, Sweden 14 Growth with number of stages Deadline Miss Rate Analysis of Applications with Stochastic Task Execution Times Sorin Manolache, Petru Eles, Zebo Peng – Linkoping University, Sweden 15 Accuracy Accuracy vs analysis complexity compared to the exact approach Stages 2 3 4 5 Relative error 8.7% 4.1% 1.04% 0.4% Deadline Miss Rate Analysis of Applications with Stochastic Task Execution Times Sorin Manolache, Petru Eles, Zebo Peng – Linkoping University, Sweden 16 Independence Assumption-Based Approximation Independence assumption-based Faster and approximate analysis for multiprocessor systems [ICCAD 2002] However it is still too slow to be plugged into an optimization loop Analysis complexity is reduced by two means: Task start and finish times are approximated with discrete values Two types of dependencies between some random variables are neglected Deadline Miss Rate Analysis of Applications with Stochastic Task Execution Times Sorin Manolache, Petru Eles, Zebo Peng – Linkoping University, Sweden 18 Independence of predecessors A Z Y X Z Y X Z Y X P(X>max(Y, Z)) = P(X>Y) P(X>Z) Deadline Miss Rate Analysis of Applications with Stochastic Task Execution Times Sorin Manolache, Petru Eles, Zebo Peng – Linkoping University, Sweden 19 Load-arrival time independence A C B A B C Time P(LC(t)) = P(LC(t)|AC<t) Deadline Miss Rate Analysis of Applications with Stochastic Task Execution Times Sorin Manolache, Petru Eles, Zebo Peng – Linkoping University, Sweden 20 Approximation effects Deadline Miss Rate Analysis of Applications with Stochastic Task Execution Times Sorin Manolache, Petru Eles, Zebo Peng – Linkoping University, Sweden 21 Experimental results Deadline Miss Rate Analysis of Applications with Stochastic Task Execution Times Sorin Manolache, Petru Eles, Zebo Peng – Linkoping University, Sweden 22 22 Conclusions Two approaches for obtaining approximations of deadline miss ratios Based on the approximation of the ETPDF by Coxian distributions Efficient scheme to store the underlying stochastic process and to construct it on the fly Based on independence assumptions among random variables Both approaches provide the possibility to trade analysis speed and memory demand for analysis accuracy Deadline Miss Rate Analysis of Applications with Stochastic Task Execution Times Sorin Manolache, Petru Eles, Zebo Peng – Linkoping University, Sweden 23